What distinguishes us from other animals is some kludgy stuff on top that enables us to:
1) chain our data prediction intuitions into extended trains of thought via System 2 style reasoning
2) share our thoughts with others via language
(and probably 1 and 2 co-evolved in a way that depended on each other)
I say that the sensory data learning mechanism is the core smarts, rather than the trains-of-thought stuff, because the former is where the bulk of the computation takes place, and the latter is some relatively simple algorithms that channel that computation into useful work.
(Analogous to the kinds of algorithms Ought and others are building to coax GPT-3 into doing useful work.)
Modern ML systems are doing basically the same thing as the predictive processing / System 1 / core smarts in our brains.
The details are different, but if you zoom out a bit, it’s basically the same algorithm. And the natural and artificial systems are able to successfully model, compress, and predict data for basically the same reasons.
AI systems will get closer to being able to match the full range of abilities of humans (and then exceed them) due to progress both on:
1) improved intuition / data compression and prediction that comes from training bigger ML models for longer on more data, and
2) better algorithms for directing those smarts into useful work.
This means that basically no new fundamental insights are needed to get to AGI / TAI. It’ll just be a bunch of iterative work to scale ML models, and productively direct their outputs.
So the path from here looks pretty continuous, though there could be some jumpy parts, especially if some people are unusually clever (or brash) with the better algorithms (for making use of model outputs) part.
I’m curious if others agree with these claims. And if not, which parts seem most wrong?
This sounds reasonable to me but I find myself wondering just how simple the “relatively simple algorithms” for channelling predictive processing are. Could you say a bit more about what the iterative path to progress looks like for improving those algorithms? (The part that seems the most wrong in what you wrote is that there are no new fundamental insights needed to improve these algorithms.)
After reading through some of the recent discussions on AI progress, I decided to sketch out my current take on where AI is and is going.
Hypotheses:
The core smarts in our brains is a process that does self-supervised learning on sensory data.
We share this smarts with animals.
What distinguishes us from other animals is some kludgy stuff on top that enables us to:
1) chain our data prediction intuitions into extended trains of thought via System 2 style reasoning
2) share our thoughts with others via language
(and probably 1 and 2 co-evolved in a way that depended on each other)
I say that the sensory data learning mechanism is the core smarts, rather than the trains-of-thought stuff, because the former is where the bulk of the computation takes place, and the latter is some relatively simple algorithms that channel that computation into useful work.
(Analogous to the kinds of algorithms Ought and others are building to coax GPT-3 into doing useful work.)
Modern ML systems are doing basically the same thing as the predictive processing / System 1 / core smarts in our brains.
The details are different, but if you zoom out a bit, it’s basically the same algorithm. And the natural and artificial systems are able to successfully model, compress, and predict data for basically the same reasons.
AI systems will get closer to being able to match the full range of abilities of humans (and then exceed them) due to progress both on:
1) improved intuition / data compression and prediction that comes from training bigger ML models for longer on more data, and
2) better algorithms for directing those smarts into useful work.
This means that basically no new fundamental insights are needed to get to AGI / TAI. It’ll just be a bunch of iterative work to scale ML models, and productively direct their outputs.
So the path from here looks pretty continuous, though there could be some jumpy parts, especially if some people are unusually clever (or brash) with the better algorithms (for making use of model outputs) part.
I’m curious if others agree with these claims. And if not, which parts seem most wrong?
This sounds reasonable to me but I find myself wondering just how simple the “relatively simple algorithms” for channelling predictive processing are. Could you say a bit more about what the iterative path to progress looks like for improving those algorithms? (The part that seems the most wrong in what you wrote is that there are no new fundamental insights needed to improve these algorithms.)